An Introduction To Football Modelling at Smartodds by Robert Johnson

An introduction to football modelling at Smartodds Robert Johnson
An introduction to football modelling at Smartodds Oxford SIAM Conference 2011

Robert Johnson
Smartodds Ltd
February 9, 2011
Introduction
Introduction to Smartodds Practical example: building a football model
What is Smartodds about?

Smartodds provides statistical research and sports modelling in the betting sector

Smartodds provides statistical research and sports modelling in the betting sector Quant team research and implement the sports models

Smartodds provides statistical research and sports modelling in the betting sector Quant team research and implement the sports models Primary focus is on Football, however we also model Basketball, Baseball, American Football, Ice Hockey and Tennis

Smartodds provides statistical research and sports modelling in the betting sector Quant team research and implement the sports models Primary focus is on Football, however we also model Basketball, Baseball, American Football, Ice Hockey and Tennis Wide range of interesting problems to work on

Smartodds provides statistical research and sports modelling in the betting sector Quant team research and implement the sports models Primary focus is on Football, however we also model Basketball, Baseball, American Football, Ice Hockey and Tennis Wide range of interesting problems to work on Actively recruiting!
Building a Football model

Suppose we decide to build a football model for the English football leagues

Suppose we decide to build a football model for the English football leagues Here we model the divisions Premier League, Championship, League 1 and League 2

Suppose we decide to build a football model for the English football leagues Here we model the divisions Premier League, Championship, League 1 and League 2 There are 92 teams in total to model

Suppose we decide to build a football model for the English football leagues Here we model the divisions Premier League, Championship, League 1 and League 2 There are 92 teams in total to model We want to predict the probability of team A winning against team B where team A and team B could be from any of the 4 leagues
Literature review
Maher (1982) assumed independent Poisson distributions for home and away goals
Literature review
Means based on each teams past performance
Literature review
Dixon and Coles (1997) took this idea further by accounting for uctuations in performance of individual teams and estimation between leagues
Literature review
Dixon and Coles (1997) took this idea further by accounting for uctuations in performance of individual teams and estimation between leagues Dixon and Robinson (1998) modelled the scores during a game as a two-dimensional birth process
Model formulation
Assume that home and away goals follow a Poisson distribution Pr (x goals) = Pr (y goals) = x e x! y e y!
Model formulation
Assume that home and away goals follow a Poisson distribution Pr (x goals) = Pr (y goals) = x e x! y e y!
To estimate the probabilities of x and y goals we need and
Model 1: Mean goals

Assume that home and away teams are expected to score the same number of goals
Model 1: Mean goals

Assume that home and away teams are expected to score the same number of goals Take average goals scored in a game in England as 2.56 and divide by two = 1.28 = 1.28
Model 1: Mean goals

Assume that home and away teams are expected to score the same number of goals Take average goals scored in a game in England as 2.56 and divide by two = 1.28 = 1.28 However we may believe that there is some advantage associated with playing at home
Model 2: Home Advantage

Include a term to take account of home advantage = =
Model 2: Home Advantage

Include a term to take account of home advantage = = is the common mean and represents the home advantage
Model 2: Home Advantage (Cont)

Mean goals scored by the away team in the four leagues we model English Leagues is 1.10 giving = 1.10

Mean goals scored by the away team in the four leagues we model English Leagues is 1.10 giving = 1.10 This implies mean goals scored by the home team are 2.56 1.10 = 1.46

Mean goals scored by the away team in the four leagues we model English Leagues is 1.10 giving = 1.10 This implies mean goals scored by the home team are 2.56 1.10 = 1.46 Using the above we can estimate as = 1.46/1.10 = 1.33
Model 3: Team Strengths

Previous attempts assumed all teams of equal strength

Previous attempts assumed all teams of equal strength Can add team strength parameters for each team

Previous attempts assumed all teams of equal strength Can add team strength parameters for each team Better teams score more goals. Give each team an attack parameter denoted

Previous attempts assumed all teams of equal strength Can add team strength parameters for each team Better teams score more goals. Give each team an attack parameter denoted Better teams concede fewer goals. Give each team a defence parameter denoted
Model 3: Team Strengths (Cont)

Write and in terms of the attack and defence parameters of the home and away teams, which we denote by i and j , giving = i j = j i
Model 3: Team Strengths (Cont)

Write and in terms of the attack and defence parameters of the home and away teams, which we denote by i and j , giving = i j = j i The model is overparameterised, so we apply the constraints 1 n
n
i = 1,
i =1
1 n
i = 1.
i =1
Model 3: Pseudolikelihood
The pseudolikelihood for this model is: L(, , i , i ; i = 1, . . . , n) =

yk (t tk ) k {exp(k )x k exp(k )k } k

() is an exponential downweighting function, which allows us to place less weight on older games

() is an exponential downweighting function, which allows us to place less weight on older games Other downweighting functions could be used
Estimation techniques
Obtaining the parameter estimates is not straightforward
Obtaining the parameter estimates is not straightforward In this example we have 186 parameters to estimate
Obtaining the parameter estimates is not straightforward In this example we have 186 parameters to estimate Various optimisation techniques could be used to obtain parameter estimates (numerical maximisation of the likelihood function, MCMC)
Obtaining the parameter estimates is not straightforward In this example we have 186 parameters to estimate Various optimisation techniques could be used to obtain parameter estimates (numerical maximisation of the likelihood function, MCMC) High dimensional problems may also require more sophisticated computing solutions (MPI)
Parameter estimates
These are Smartodds current estimates of the attack and defence parameters of the top 6 teams in the Premier League Team Chelsea Man Utd Arsenal Man City Tottenham Liverpool Attack Parameter 3.15 3.08 2.84 2.44 2.22 2.12 Defence Parameter 0.34 0.35 0.37 0.42 0.44 0.39
Predicting outcomes
Suppose Man Utd are playing at home to Man City
Predicting outcomes
Suppose Man Utd are playing at home to Man City Using the parameter estimates we get = 1.10 1.33 3.08 0.42 = 1.89 = 1.10 2.44 0.35 = 0.94
Predicting outcomes
Suppose Man Utd are playing at home to Man City Using the parameter estimates we get = 1.10 1.33 3.08 0.42 = 1.89 = 1.10 2.44 0.35 = 0.94 We can use and to obtain the probability of Man Utd winning the match
Predicting outcomes (Cont)

The probability of a specic score is given as follows x e y e Pr (x , y ) = x! y!

The probability of a specic score is given as follows x e y e Pr (x , y ) = x! y! So the probability of the score, Man Utd 2 Man City 1, is Pr (2, 1) = 1.892 e 1.89 0.941 e 0.94 = 0.099 2! 1!

Obtain the probability matrix of all possible scores 0 1 2 3 4 . . . 0 0.059 0.055 0.026 0.008 0.002 . . . 1 0.112 0.105 0.049 0.015 0.004 . . . 2 0.105 0.099 0.047 0.015 0.003 . . . 3 0.066 0.062 0.029 0.009 0.002 . . . 4 0.031 0.029 0.014 0.004 0.001 . . . ... ... ... ... ... ... .. .

Sum over all events where home goals are greater than away goals 0 0.059 0.055 0.026 0.008 0.002 . . . 1 0.112 0.105 0.049 0.015 0.004 . . . 2 0.105 0.099 0.047 0.015 0.003 . . . 3 0.066 0.062 0.029 0.009 0.002 . . . 4 0.031 0.029 0.014 0.004 0.001 . . . ... ... ... ... ... ... .. .
0 1 2 3 4 . . .

Giving the probability that Man Utd win at home to Man City as 59.6% 0 0.059 0.055 0.026 0.008 0.002 . . . 1 0.112 0.105 0.049 0.015 0.004 . . . 2 0.105 0.099 0.047 0.015 0.003 . . . 3 0.066 0.062 0.029 0.009 0.002 . . . 4 0.031 0.029 0.014 0.004 0.001 . . . ... ... ... ... ... ... .. .
0 1 2 3 4 . . .
Practical issues
Betfairs odds imply Man Utd has a 63% chance of winning the game, potentially leaving value for a bet on Man City. However, should we bet?
Practical issues
Betfairs odds imply Man Utd has a 63% chance of winning the game, potentially leaving value for a bet on Man City. However, should we bet? These models take into account no external information about match circumstances
Practical issues
Injuries
Practical issues
Injuries Motivation
Practical issues
Injuries Motivation Fatigue
Practical issues
Injuries Motivation Fatigue Newly signed players
Practical issues
Injuries Motivation Fatigue Newly signed players
So betting o a mathematical model would be dangerous!
Shortcomings of the model

If we compare the expected full-time scores under the model with the observed scores, we nd our modelling assumptions dont hold

Goals dont have a Poisson distribution

Goals dont have a Poisson distribution Goals scored by the home and away teams arent independent

Dixon and Coles corrected for this by modifying the predicted distribution to increase probability of draws and 0-1 and 1-0 scores

Dixon and Coles corrected for this by modifying the predicted distribution to increase probability of draws and 0-1 and 1-0 scores However this isnt entirely satisfactory would be better to model what is happening directly
Goal time distribution

Density 0.05 0.00
0
0.01
0.02
0.03
0.04
10
20
30
40
50
60
70
80
90
Goal time (mins)
Goals in injury time at the end of each half are recorded as 45 / 90 min goals

Density 0.05 0.00
0
0.01
0.02
0.03
0.04
10
20
30
40
50
60
70
80
90
Goal time (mins)
Goals in injury time at the end of each half are recorded as 45 / 90 min goals Goal rate steadily increases over the course of the game

Density 0.05 0.00
0
0.01
0.02
0.03
0.04
10
20
30
40
50
60
70
80
90
Goal time (mins)
Goals in injury time at the end of each half are recorded as 45 / 90 min goals Goal rate steadily increases over the course of the game Notice the spikes every 5 minutes in the second half - due to rounding?
Dixon and Robinsons model

If we assume that the goal scoring processes for the home and away teams are independent homogeneous Poisson processes then our model reduces to the full time model discussed previously.
Dixon and Robinsons model

If we assume that the goal scoring processes for the home and away teams are independent homogeneous Poisson processes then our model reduces to the full time model discussed previously. For match k between teams i and j k (t ) = k = i j k (t ) = k = j i
Dixon and Robinsons model (continued)

Three changes:

Three changes:
1
Goal-scoring rate dependent on the current score

Three changes:
1 2
Goal-scoring rate dependent on the current score Modelling of injury time

Three changes:
1 2 3
Goal-scoring rate dependent on the current score Modelling of injury time Increasing goal-scoring intensity through the game (due to tiredness of players)
(1) Goal-scoring rate dependent on current score

Assume that home and away scoring processes are independent Poisson processes

Assume that home and away scoring processes are independent Poisson processes Scoring rates are piecewise constant

Home and away intensities are constant until a goal is scored and only change at these times

Denote xy and xy as parameters determining the scoring rates when the score is (x ,y )

Denote xy and xy as parameters determining the scoring rates when the score is (x ,y ) Scoring rates are now k (t ) = xy k

Denote xy and xy as parameters determining the scoring rates when the score is (x ,y ) Scoring rates are now k (t ) = xy k and k (t ) = xy k
Estimates of (x , y ) and (x , y )
(0, 0) = 1 (0, 0) = 1
(0, 0) (0, 0) (1, 0) (1, 0)
=1 =1 = 0.88 = 1.35
(0, 0) (0, 0) (1, 0) (1, 0) (0, 1) (0, 1)
=1 =1 = 0.88 = 1.35 = 1.10 = 1.07
(2) Increase the scoring rate during injury time

Goals scored during injury time are recorded as having occurred at either 45 or 90 minutes.

Goals scored during injury time are recorded as having occurred at either 45 or 90 minutes. Dene two new parameters 1 and 2 to model injury time.

Goals scored during injury time are recorded as having occurred at either 45 or 90 minutes. Dene two new parameters 1 and 2 to model injury time. The adjusted scoring rates are 1 xy k t (44, 45]mins, k (t ) = 2 xy k t (89, 90]mins, xy k otherwise

Goals scored during injury time are recorded as having occurred at either 45 or 90 minutes. Dene two new parameters 1 and 2 to model injury time. The adjusted scoring rates are 1 xy k t (44, 45]mins, k (t ) = 2 xy k t (89, 90]mins, xy k otherwise and similarly for k (t )
(3) Increasing goal-scoring intensity

Allow the scoring intensities to increase over time

Allow the scoring intensities to increase over time Model scoring rates as time inhomogeneous Poisson processes with a linear rate of increase

Allow the scoring intensities to increase over time Model scoring rates as time inhomogeneous Poisson processes with a linear rate of increase Replace k (t ) and k (t ) with k ( t ) = k ( t ) + 1 t , k (t ) = k (t ) + 2 t

Allow the scoring intensities to increase over time Model scoring rates as time inhomogeneous Poisson processes with a linear rate of increase Replace k (t ) and k (t ) with k ( t ) = k ( t ) + 1 t , k (t ) = k (t ) + 2 t 1 and 2 could be constrained to be positive to ensure that the hazard functions above are constrained to always be positive, but in practice this is not neccessary

Allow the scoring intensities to increase over time Model scoring rates as time inhomogeneous Poisson processes with a linear rate of increase Replace k (t ) and k (t ) with k ( t ) = k ( t ) + 1 t , k (t ) = k (t ) + 2 t 1 and 2 could be constrained to be positive to ensure that the hazard functions above are constrained to always be positive, but in practice this is not neccessary Scoring rates are estimated to be about 75% higher at the end of the game then at the start of the game.
Model usage
This in-running model can be useful in its own right (for deriving in-running prices)
Model usage
This in-running model can be useful in its own right (for deriving in-running prices) Also explains the home/away dependencies and non-Poisson pdfs observed in the data
Summary
The Dixon-Coles model is a simple and robust full-time score model, but not all of its assumptions are met
Summary
The Dixon-Coles model is a simple and robust full-time score model, but not all of its assumptions are met A continuous time model such as the Dixon-Robinson model can model dependencies between home and away scoring rates
Summary
The Dixon-Coles model is a simple and robust full-time score model, but not all of its assumptions are met A continuous time model such as the Dixon-Robinson model can model dependencies between home and away scoring rates Mathematical models cannot model team news (unless this is incorporated into the model somehow)
Summary
The Dixon-Coles model is a simple and robust full-time score model, but not all of its assumptions are met A continuous time model such as the Dixon-Robinson model can model dependencies between home and away scoring rates Mathematical models cannot model team news (unless this is incorporated into the model somehow) These models can be extended to other sports by changing the distributions, eg
Summary
Normal distribution for American Football
Summary
Normal distribution for American Football Negative binomial for baseball
References
M.J. Maher, 1982, Modelling association football scores, Statist. Neerland., 36, 109-1188
References
M.J. Maher, 1982, Modelling association football scores, Statist. Neerland., 36, 109-1188 M. Dixon and S.G. Coles, 1997. Modelling Association Football Scores and Ineciencies in the Football Betting Market. Applied Statistics, 46(2), 265-280
References
M.J. Maher, 1982, Modelling association football scores, Statist. Neerland., 36, 109-1188 M. Dixon and S.G. Coles, 1997. Modelling Association Football Scores and Ineciencies in the Football Betting Market. Applied Statistics, 46(2), 265-280 M. Dixon and M. Robinson, 1998. A birth process model for association football matches. JRSS D, 47(3), 523-538
Interested?
If you are interested in sports modelling and possess the following skills:
Interested?
Post graduate qualication (at least MMath / MSc, PhD. preferred) in mathematics, statistics or another subject with considerable mathematical content
Interested?
Post graduate qualication (at least MMath / MSc, PhD. preferred) in mathematics, statistics or another subject with considerable mathematical content Experience in developing and implementing mathematical / statistical models
Interested?
Post graduate qualication (at least MMath / MSc, PhD. preferred) in mathematics, statistics or another subject with considerable mathematical content Experience in developing and implementing mathematical / statistical models Experience of computer programming (preferably in C++, C, R or Python)
Interested?
Post graduate qualication (at least MMath / MSc, PhD. preferred) in mathematics, statistics or another subject with considerable mathematical content Experience in developing and implementing mathematical / statistical models Experience of computer programming (preferably in C++, C, R or Python) Enthusiasm, self-motivation and the ability to work under pressure to strict deadlines
Interested?
Then email us at careers@smartodds.co.uk
Interested?
Then email us at careers@smartodds.co.uk For more information see our website: http://www.smartodds.co.uk

An Introduction To Football Modelling at Smartodds by Robert Johnson

Caricato da

Informazioni sul documento

Titolo originale

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

An Introduction To Football Modelling at Smartodds by Robert Johnson

Caricato da

Copyright:

Formati disponibili

An introduction to football modelling at Smartodds Robert Johnson

An introduction to football modelling at Smartodds Oxford SIAM Conference 2011

Introduction to Smartodds Practical example: building a football model

What is Smartodds about?

What is Smartodds about?

What is Smartodds about?

What is Smartodds about?

What is Smartodds about?

Building a Football model

Building a Football model

Building a Football model

Building a Football model

To estimate the probabilities of x and y goals we need and

Model 1: Mean goals

Model 1: Mean goals

Model 1: Mean goals

Model 2: Home Advantage

Include a term to take account of home advantage = =

Model 2: Home Advantage

Model 2: Home Advantage (Cont)

Model 2: Home Advantage (Cont)

Model 2: Home Advantage (Cont)

Model 3: Team Strengths

Previous attempts assumed all teams of equal strength

Model 3: Team Strengths

Model 3: Team Strengths

Model 3: Team Strengths

Model 3: Team Strengths (Cont)

Model 3: Team Strengths (Cont)

The pseudolikelihood for this model is: L(, , i , i ; i = 1, . . . , n) =

The pseudolikelihood for this model is: L(, , i , i ; i = 1, . . . , n) =

The pseudolikelihood for this model is: L(, , i , i ; i = 1, . . . , n) =

Obtaining the parameter estimates is not straightforward

Suppose Man Utd are playing at home to Man City

Predicting outcomes (Cont)

The probability of a specic score is given as follows x e y e Pr (x , y ) = x! y!

Predicting outcomes (Cont)

Predicting outcomes (Cont)

Predicting outcomes (Cont)

Predicting outcomes (Cont)

So betting o a mathematical model would be dangerous!

Shortcomings of the model

Shortcomings of the model

Shortcomings of the model

Shortcomings of the model

Shortcomings of the model

Goal time distribution

Goal time (mins)

Goal time distribution

Goal time (mins)

Goal time distribution

Goal time (mins)

Dixon and Robinsons model

Dixon and Robinsons model

Dixon and Robinsons model (continued)

Dixon and Robinsons model (continued)

Goal-scoring rate dependent on the current score

Dixon and Robinsons model (continued)

Goal-scoring rate dependent on the current score Modelling of injury time

Dixon and Robinsons model (continued)

(1) Goal-scoring rate dependent on current score

(1) Goal-scoring rate dependent on current score

(1) Goal-scoring rate dependent on current score